The kinds of information used in on-line syntactic processing decisions has long been a source of controversy. We are investigating a computational model of minimalist syntactic theory (e.g. Weinberg and Berwick's paper of that title at the 1997 Conference on Computational Psycholinguistics), and recently have begun looking into the role of plausibility in on-line processing using eye-tracking methods.
Language theorists have long noted that language places constraints on the relationships between things being talked about, constraints that are sometimes violated even when the surface form of a sentence is perfectly valid. These constraints are often referred to as selectional preferences: a word is said to ``select for'' the kinds of things with which it can be associated. Chomsky's famous sentence ``Colorless green ideas sleep furiously'' violates a number of these preferences --- for example, the verb 'sleep' selects for a subject that is animate, which 'ideas' are not. This research concerns the computational modeling of selectional constraints, and the use of such models in (1) making predictions about linguistic behavior, and (2) performing computational tasks such as word sense disambiguation. Some relevant publications can be found on Philip Resnik's list of publications.
Evaluating semantic similarity and relatedness using network representations is a problem with a long history in artificial intelligence and psychology, dating back to the spreading activation approach of Quillian (1968) and Collins and Loftus (1975). Our work is investigating the definition of a taxonomic similarity measure based on the notion of information content, evaluated by how well it approximates human similarity ratings, and the application of that and related similarity measures in word sense disambiguation.